A 128GB DDR5 Kit That Bottomed Out at $329 Now Lists at $3,399. The Self-Hosting Math Every Solo Operator Did Last Year Is Dead.
A DDR5-6400 128GB kit that bottomed out at $329 is now listed at $3,399. That is ten times its own record low. Across the broader market, tracked memory prices are up roughly 500% over twelve months, and a 64GB DDR5-5600 kit that ran under $200 last summer is now over $1,100.
Two weeks ago I wrote that RAM was up around 89% on the year and that your next dev machine or VPS bill was not exempt. That number is now badly out of date, and the new one is large enough that it does not just change the headline. It changes the recommendation.
The numbers, and where they came from
The figures come from retail price tracking, so treat them as list prices at specific vendors rather than as a clean index. The pattern across the reported kits is consistent enough to act on:
- 128GB DDR5-6400: from a $329 low to $3,399
- 64GB DDR5-5600: from roughly $191 to over $1,100
- 32GB DDR5-6000: from roughly $222 to roughly $1,272
The stated driver is hyperscale AI customers reserving large portions of future DRAM capacity. I want to be careful about the causal claim, because "AI datacenters caused this" is the framing every outlet reached for and none of them established it rigorously. What is well supported is that hyperscalers are booking forward capacity and that supply to the retail channel has tightened. Whether that is the whole explanation, versus one input alongside fab allocation decisions and the usual memory cycle, is not something the reporting settles.
The direction is not in dispute. The magnitude is what is new.
What this does to the local inference argument
Here is the case that got popular, roughly, from late 2025 onward. Capable models started fitting on hardware a person could own. Buy a box with a lot of memory, run a 30B-class model locally, stop paying per-token API bills, own your inference. Amortised over eighteen months, the numbers worked.
The numbers worked because RAM was cheap. That was the load-bearing assumption, and nobody wrote it down because it had been true for two decades.
Run the arithmetic at current prices. If the memory alone in a 128GB build is $3,399, you need to displace a lot of API spend to break even before the hardware is obsolete. At the promotional frontier-model rates going around this month, you would need to be burning tokens at a rate most solo operators simply do not hit. My own inference spend across every project has never come close, and I am on the heavier end of what a one-person shop does.
There is a second problem specific to this moment. The break-even case for owned hardware depends on holding it long enough. But the hardware requirements for local models have been moving fast, and buying at a 10x price peak is the worst possible entry point for an asset you need to hold for three years to justify.
The part that lags: VPS and dedicated servers
Retail DRAM prices move first. Hosting prices move later, and differently.
Providers buy on contracts and hold inventory, so a spike does not pass through immediately. What tends to happen instead is that existing plans keep their prices while quietly getting worse: the RAM-heavy tiers stop being offered, the promotional pricing on new signups disappears first, and the generous memory allocations that made a plan attractive get trimmed on the next generation of hardware.
So the thing to watch is not your renewal invoice. It is whether the plan you are on is still on the provider's public pricing page in six months. If your workload depends on a memory-heavy tier at a price that looks too good, the risk is not a price rise, it is a migration you did not schedule.
I do not have hard data on how much of the current spike has passed through to hosting yet, and I would be sceptical of anyone claiming a precise figure this early. The mechanism is well understood, the timing is not.
What I'd actually do
If you were planning to buy memory this quarter for a local inference box, do not. This is the clearest call in the post. You are buying at ten times a record low into a market with a supply explanation that could resolve, and renting inference has never been cheaper relative to owning it.
If you already own the hardware, none of this touches you and you should ignore the entire conversation. Sunk cost works in your favour here for once. Your amortisation schedule is unaffected by what a new kit costs today, and if anything your box just became a more valuable asset than it was in January.
If you need memory for actual work rather than for inference, look at DDR4 and the used market before the new DDR5 channel. Older generations and secondhand supply do not track the leading-edge spike closely, and for a development machine that is not running frontier models the performance difference is frequently not what is limiting you.
And if you are on a memory-heavy hosting plan, screenshot the current pricing page and set a reminder for three months. Not because the price will change, but so you can tell whether the tier still exists.
Where I might be wrong
The strongest counter is that these are retail list prices at a moment of acute shortage, and retail list is the noisiest possible signal. Spot and contract pricing behave differently, availability varies by region, and "up 500%" compares against a record low, which is a comparison chosen to be dramatic. Against a three-year average the increase is real but less spectacular.
There is also a timing argument against me. If the shortage resolves and prices fall back over the next year, the person who waited saved money, but the person who bought and started running local models a year earlier got a year of compounding familiarity with a stack that is going to matter. That is not nothing, and I do not weight learning time at zero.
The narrow claim I will defend: for a solo operator running a normal workload, buying $3,399 of DDR5 to avoid API bills does not survive a spreadsheet at today's prices. Everything else in this post is a judgement call about a market that has already made two decades of people look silly for predicting it.
Author
Lukas
@lukcombinator